SMT-GGCN: Self-Supervised Multi-Task Learning for Generative Graph Convolutional Network

Deep learning on graphs has shown to be powerful in modeling graph-structured data and learning representations for graphs, which involves aggregating information from graph neighborhoods to derive effective node representations. Although graph neural networks (GNNs) can be trained from scratch, pretraining GNNs to learn transferable knowledge for downstream tasks has been used recently to improve performance. However, using pre-training and fine-tuning to model each subtask can sometimes ignore the relationships, conflicts, and constraints between tasks, resulting in worse performance. In this paper, we propose a novel framework called Self-supervised Multi-Task Learning for Generative Graph Convolutional Network (SMT-GGCN) to enhance graph-based neural network models. SMT-GGCN combines self-supervised learning pretext tasks with supervised learning to train Graph Convolutional Networks (GCN) for graph structure understanding. We demonstrate that self-supervised multi-task learning can improve GCN’s generalization ability and enhance feature learning.